remove @inbounds

[simeonschaub]

In my benchmarks, Julia 1.7 seems to be smart enough to eliminate the bounds check on its own:

julia> function _sum(A)
           s = 0.0
           for a in A
               s += a
           end
           s
       end
_sum (generic function with 1 method)

julia> function _sum2(A)
           s = 0.0
           @inbounds @simd for a in A
               s += a
           end
           s
       end
_sum2 (generic function with 1 method)

julia> function _sum3(A)
           s = 0.0
           @simd for a in A
               s += a
           end
           s
       end
_sum3 (generic function with 1 method)

julia> using BenchmarkTools

julia> a = rand(10^7);

julia> @benchmark _sum($a)
BenchmarkTools.Trial: 425 samples with 1 evaluation.
 Range (min … max):  10.595 ms …  13.903 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     11.753 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   11.770 ms ± 177.003 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

                                             █▂▁
  ▅▁▁▁▁▁▁▁▁▁▄▁▁▁▁▁▄▁▁▁▁▁▁▁▁▁▁▁▄▁▁▁▁▁▁▁▁▁▁▁▁▁▁███▅▅▅▄▄▄▁▁▄▄▆▁▁▄ ▆
  10.6 ms       Histogram: log(frequency) by time      12.2 ms <

 Memory estimate: 0 bytes, allocs estimate: 0.

julia> @benchmark _sum2($a)
BenchmarkTools.Trial: 1024 samples with 1 evaluation.
 Range (min … max):  4.323 ms …   8.279 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     4.850 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   4.877 ms ± 207.137 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

                       █▂▄▅▂▁▁▂▃
  ▂▁▂▂▁▂▁▁▂▁▂▂▂▂▂▁▂▁▂▁██████████▇▆▄▃▃▃▂▂▂▂▂▃▂▂▂▂▂▁▁▁▁▁▂▁▁▁▁▁▂ ▃
  4.32 ms         Histogram: frequency by time        5.55 ms <

 Memory estimate: 0 bytes, allocs estimate: 0.

julia> @benchmark _sum3($a)
BenchmarkTools.Trial: 988 samples with 1 evaluation.
 Range (min … max):  4.117 ms …   8.226 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     4.935 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   5.049 ms ± 579.887 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

              ▄█▂
  ▃▃▃▂▁▃▃▃▃▄▅▆███▆▆▅▅▅▄▃▃▂▂▂▂▁▂▂▂▁▁▂▁▁▁▂▂▂▁▂▂▁▁▁▁▁▂▁▂▂▂▂▂▂▂▂▂ ▃
  4.12 ms         Histogram: frequency by time        7.63 ms <

 Memory estimate: 0 bytes, allocs estimate: 0.

[simeonschaub]

There is no array accessing (e.g. A[2]) in this piece of code, so _sum2 and _sum3 are essentially the same and @inbounds is actually redundant in _sum2.

Just because you don't see it doesn't mean this code doesn't end up calling A[i]. :) It might have not been the best example though, because that call is already annotated as @inbounds: https://github.com/JuliaLang/julia/blob/2d6a84ecedc02a58d0a4f64d898ae5124ad35c2b/base/array.jl#L895.

The programming language should allow users to control this instead of being 'smart' enough to eliminate the bounds check on its own.

I definitely disagree with this. In the majority of cases, Julia's compiler should have all the information it needs to eliminate bounds check. @inbounds should really be thought of more as a bandaid for cases where the compiler isn't yet smart enough. It's very easy to use incorrectly and the problem is that you might not just get a segfault immediately, instead out-of-bounds accesses can also lead to quite memory corruption and hard-to-track bugs down the road. That's why ideally, we'd like to just get rid of @inbounds alltogether.

Observe for example that even with explicit indexing, LLVM can still eliminate bounds checks (unfortunately only on Julia master right now):

julia> function _sum(A)
           s = 0.0
           for i in eachindex(A)
               s += A[i]
           end
           s
       end
_sum (generic function with 1 method)

julia> function _sum2(A)
           s = 0.0
           @inbounds @simd for i in eachindex(A)
               s += A[i]
           end
           s
       end
_sum2 (generic function with 1 method)

julia> function _sum3(A)
           s = 0.0
           @simd for i in eachindex(A)
               s += A[i]
           end
           s
       end
_sum3 (generic function with 1 method)

julia> @benchmark _sum($a)
BenchmarkTools.Trial: 591 samples with 1 evaluation.
 Range (min … max):  8.116 ms …   9.745 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     8.422 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   8.447 ms ± 134.948 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

                 ▃▁▁▇▆█▄█▄▅▄   ▃▂                              
  ▂▁▁▂▁▁▃▂▃▃▅▂▇█▇█████████████████▅▃▆▆▅▄▃▃▃▄▂▁▁▁▃▁▂▁▃▃▂▂▁▂▁▁▃ ▄
  8.12 ms         Histogram: frequency by time        8.91 ms <

 Memory estimate: 0 bytes, allocs estimate: 0.

julia> @benchmark _sum2($a)
BenchmarkTools.Trial: 1174 samples with 1 evaluation.
 Range (min … max):  3.880 ms …   5.134 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     4.109 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   4.241 ms ± 279.201 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

      ▁▄▆▆█▆▇▄▁▁                                               
  ▁▂▄▅██████████▆▆▅▄▃▄▂▃▂▄▃▄▃▅▄▂▄▃▄▄▃▄▄▄▄▃▃▃▄▄▃▃▄▄▃▂▃▄▃▂▂▃▃▂▂ ▃
  3.88 ms         Histogram: frequency by time        4.92 ms <

 Memory estimate: 0 bytes, allocs estimate: 0.

julia> @benchmark _sum3($a)
BenchmarkTools.Trial: 1188 samples with 1 evaluation.
 Range (min … max):  3.781 ms …   5.362 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     4.064 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   4.191 ms ± 286.753 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

         ▂▂▅█▅▃▆▃                                              
  ▂▃▃▄▄▇▇██████████▆▄▅▄▃▄▅▄▃▄▄▄▃▃▅▄▄▄▃▅▄▄▅▃▄▃▄▅▆▅▅▄▃▃▃▃▃▂▂▁▂▂ ▄
  3.78 ms         Histogram: frequency by time        4.94 ms <

 Memory estimate: 0 bytes, allocs estimate: 0.